CSST Slitless Spectra: Target Detection and Classification with YOLO

Yingying Zhou, Chao Liu, Hao Tian, Xin Zhang, Nan Li · 2025

Read CSST Slitless Spectra: Target Detection and Classification with YOLO by Yingying Zhou, Chao Liu, Hao Tian, Xin Zhang, Nan Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-24087.txt.

Introduction

Yingying Zhou and four co-authors apply object detection models to a hard task in space astronomy: reading slitless spectra. The China Space Station Telescope will take images where spectra overlap and their place is uncertain. Old step by step methods build up errors as they extract and sort these traces.

Their framework uses YOLO models to find, sort, and study traces directly from raw two dimensional images. YOLOv5 finds both compact and stretched traces even in crowded fields, while YOLO11 adds source sorting and rough estimates of traits like redshift. The system works fast on large images and can spot quasars, showing an end to end path with little manual work.

Catalog record

CREATORYingying Zhou, Chao Liu, Hao Tian, Xin Zhang, Nan Li
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE98 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-24087

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

CSST Slitless Spectra: Target Detection and Classification with YOLO Yingying Zhou, Chao Liu, Hao Tian, Xin Zhang, Nan Li Abstract Addressing the spatial uncertainty and spectral blending challenges in CSST slitless spectroscopy, we present a deep learning-driven, end-to-end framework based on the You Only Look Once (YOLO) models. This approach directly detects, classifies, and analyzes spectral traces from raw 2D images, bypassing traditional, error-accumulating pipelines. YOLOv5 effectively detects both compact zero-order and extended first-order traces even in highly crowded fields. Building on this, YOLO11 integrates source classification (star/galaxy) and discrete astrophysical parameter estimation (e.g., redshift bins), showcasing complete spectral trace analysis without other manual preprocessing. Our framework processes large images rapidly, learning spectral-spatial features holistically to minimize errors. We achieve high trace detection precision (YOLOv5) and demonstrate successful quasar identification and binned redshift estimation (YOLO11). This study establishes machine learning as a paradigm shift in slitless spectroscopy, unifying detection, classification, and preliminary parameter estimation in a scalable system. Future research will concentrate on direct, continuous prediction of astrophysical parameters from raw spectral traces. show][email protected] I Introduction I.1 Overview of CSST and Motivation The China Space Station Telescope (CSST) is set to serve as China’s premier astronomical facility over the coming decades, with a projected operational lifespan of ten years. Designed to survey approximately 17,500 square degrees of the sky, CSST combines high spatial resolution ( \leq 0.15 arcseconds) with a multi-channel observational framework. Its imaging capabilities span seven photometric bands (NUV, u, g, r, i, z, y), achieving a detection depth of 26 mag in the g band with 300-second exposures and an average depth of 25.5 mag in NUV, u, r, i, and z bands. Complementing these imaging surveys, CSST’s slitless spectroscopy covers three wavelength ranges: GU (255–420 nm), GV (400–650 nm), and GI (620–1000 nm), with a spectral resolution of R \lesssim 200 and broadband sensitivities reaching 22–23 mag. Ultra-deep surveys over 400 square degrees will extend these limits by an additional magnitude, enabling unprecedented studies of faint astronomical objects ( Zhan 2011 ; Zhan 2018 ; Zhan 2021 ) . CSST holds exceptional potential for resolving the Milky Way’s central bulge, where it will obtain low-resolution spectra for 10–20 million stars and multi-band photometry for 1–2 billion stars. These data will refine measurements of stellar metallicities, kinematics, and ages, offering new constraints on the properties of past merger events (e.g., progenitor mass, merger chronology) and identifying previously undetected accretion remnants. Such insights promise to advance reconstructions of the Milky Way’s hierarchical assembly history. Beyond the Galaxy, CSST’s wide-field capabilities will resolve individual stars in nearby galaxies, enabling detailed analyses of their star formation histories, metallicity distributions, and tidal interactions. By mapping stellar streams and halo substructures, the telescope will shed light on galaxy evolution mechanisms, including satellite accretion, dark matter halo dynamics, and the role of mergers in shaping disk galaxies. CSST’s deep, wide-area surveys will probe cosmological phenomena such as dark energy, large-scale structure growth, and galaxy clustering. Its photometric redshift accuracy and slitless spectroscopy will enhance measurements of weak gravitational lensing, baryon acoustic oscillations (BAO), and redshift-space distortions. These datasets will tighten constraints on cosmological parameters, including the Hubble constant and the equation of state of dark energy, while ultra-deep fields will trace early galaxy formation, bridging observations of the high-redshift universe with local galactic ecosystems. By unifying studies of the Milky Way, nearby galaxies, and the distant universe, CSST will establish a cross-scale framework for addressing fundamental questions in astrophysics, from the assembly of stellar halos to the expansion history of the cosmos. CSST employs an innovative focal plane configuration distinct from conventional telescope designs. To maximize observational throughput, its primary focal plane integrates both slitless spectroscopic detectors and multi-band imaging sensors simultaneously. While this dual-mode design enhances efficiency, it introduces a unique challenge: during initial observational phases, slitless spectroscopic data lack contemporaneous multi-color photometric references. Without overlapping imaging to anchor positional coordinates, the spatial origins of slitless spectral signals remain ambiguous. Consequently, the absence of direct source localization prevents precise extraction of spectra. CSST will produce unprecedented data volumes, presenting a daunting computational and analytical challenge. A critical issue arises in densely populated regions, such as the Milky Way’s core, where slitless spectroscopic observations inherently produce overlapping spectral traces from adjacent sources. In crowded stellar environments like the Galactic bulge, the absence of physical slits allows light from neighboring stars to blend, creating composite signals that complicate spectral disentanglement and source-specific analysis. Legacy pipelines, optimized for isolated sources, often exclude overlapping spectral regions to simplify analysis, inadvertently sacrificing critical astrophysical information. Retaining and interpreting these overlaps, however, is vital for reconstructing accurate stellar properties (e.g., chemical abundances, kinematic properties) and ensuring the scientific fidelity of CSST’s outputs. To address the dual challenges of (1) spatial uncertainties arising from non-simultaneous imaging and (2) spectral blending in crowded stellar fields, we present a deep learning-driven, end-to-end framework that decodes stellar properties directly from two-dimensional (2D) slitless spectral traces. By tackling these limitations, our methodology aims to enhance the fidelity of future CSST data products, enabling breakthroughs in stellar astrophysics, galactic archaeology, and cosmological studies. Extracting astrophysical parameters from 2D spectral images is achievable because these images preserve the full wavelength-flux relationship along the dispersion axis—mirroring the diagnostic power of 1D spectra—while simultaneously encoding spatial information along the cross-dispersion axis. For point sources (e.g., stars), the spatial axis captures the telescope’s point spread function (PSF), enabling precise modeling of how light from individual stars spreads across the detector, which helps disentangle overlapping spectra in crowded fields. For extended sources (e.g., galaxies), the spatial axis reveals spatial gradients like velocity or metallicity, providing multidimensional constraints unavailable in 1D data. With both spectral and spatial information, advanced computational techniques (e.g., machine learning, Bayesian inference) could deblend overlapping signals, resolve parameter degeneracies, and correlate spectral-spatial patterns with physical properties. This dual-axis synergy 1 1 1 It means combining spectral and position information to solve problems that one alone can’t. positions 2D spectral analysis as an indispensable tool for surveys like CSST, where the extreme stellar density of the Galactic bulge and the structural complexity of galaxies demand multidimensional diagnostics to overcome the limitations of traditional 1D methods. In this study, we focus on the framework’s foundational phase: automated detection of spectral traces in unprocessed slitless images and precise delineation of bounding boxes to isolate individual sources. This approach mitigates positional ambiguity in early-phase CSST data—where concurrent multi-band imaging is unavailable—facilitating subsequent spectral extraction and parameter derivation. Critically, we validate the model’s proficiency in identifying sources within ultra-dense regions, underscoring its potential to maximize CSST’s discovery potential in environments where traditional methods falter. I.2 The characteristic and challenges of the our dataset As previously noted, CSST data processing faces two critical challenges: (1) positional ambiguity (due to the absence of concurrent imaging for spatial anchoring) and (2) spectral blending (in crowded stellar fields). To address the first challenge, object detection techniques from computer vision—such as bounding box localization—could be adapted to identify and spatially pinpoint spectral traces directly in 2D slitless images, bypassing the need for prior imaging. Object detection aims to identify specific semantic objects within digital images and videos. These objects can belong to various categories, such as

Read the complete text